Free Flight Delay Data: Dataset Structure and Field Coverage

Each row carries the fields a delay classifier needs: scheduled and actual times, delay minutes, cancellation and diversion flags, joined to airline and airport reference tables. That completeness is why it remains the standard benchmark for delay-prediction work — you can reproduce a published result on the same 5.8 million rows, then extend the pipeline backward through the BTS on-time archive to October 1987 once it runs.

One join is worth planning from the start. The dataset's airports.csv file ships coordinates, so mapping origin-destination pairs or building distance features needs no separate geocoding step. For wider network context — equipment types, codeshare structure, carriers outside the 2015 window — OpenFlights supplies 7,698 airports, 6,162 airlines and 67,663 directional routes in plain CSV.

For the United Kingdom specifically, the UK CAA's open-government portal serves monthly and annual CSV and PDF tables of passengers and air transport movements for roughly 50-60 UK airports, with archives reaching back to 1973. These are traffic-volume tables rather than delay minutes — useful for stating how busy an airport was when delays spiked, not for measuring lateness directly.

Taken together with BTS, a free researcher gets nearly four decades of American per-flight records, over fifteen years of European en-route series and more than fifty years of UK airport throughput — without paying anyone for access.

Delay rows become far more useful once they carry geography and market context, and three free reference sets supply it.

OpenFlights provides the network backbone under the commercial delivery terms: 7,698 airports and 6,162 airlines with IATA and ICAO codes, coordinates, callsigns and codeshare flags, plus 67,663 directional routes. Plan around one caveat — the route table comes from a third-party feed frozen in June 2014, while the airport and airline files still receive periodic refreshes, so treat route-level conclusions as a 2014 snapshot.

And when fleet or ownership detail matters, the FAA registry's 316,086 active N-numbers tie tail-level questions to the same public-domain footing as the delay records themselves. Between them, these references turn a raw BTS extract into a mapped, market-aware analysis without spending anything on data access.